基于Transformer全局-局部特征融合的RGB-D显著性检测  

RGB-D SALIENCY DETECTION BASED ONTRANSFORMER GLOBAL-LOCAL FEATURE FUSION

作　　者：宋梦柯 王芸 郑元超 Song Mengke;Wang Yun;Zheng Yuanchao(College of Computer Science and Technology,Qingdao University,Qingdao 266071,Shandong,China)

机构地区：[1]青岛大学计算机科学技术学院,山东青岛266071

出　　处：《计算机应用与软件》2025年第3期176-182,共7页Computer Applications and Software

摘　　要：现有的RGB-D方法一般通过局部操作分别应用多尺度和多模态融合,但这无法捕获远程依赖性,因此对特征整体表征能力不足。针对此问题,提出一种全局-局部特征融合网络。在低层特征提取阶段,将两个分支特征直接融合;在高层特征提取阶段,将融合后特征送入Transformer编码器通过在所有位置同时整合多尺度和多模态的特征来进行充分的特征融合,获得全局特征依赖关系之后再送入主干网络提取全局—局部融合特征。同时提出双重注意力模块,用来增强两个分支特征的融合效果。在五个公开数据集上进行的实验表明,该网络在三个评价指标上均取得了较好的表现。Existing RGB-D salient object detection methods mainly usemulti-scale and multimodal fusion by local features,which cannot capture remote dependencies,so the overall characterization ability of features is insufficient.In order to solve the problem,this paper proposes a global-local feature fusion network.In the low-level feature extraction stage,the two branch features were directly fused.In the high-level feature extraction stage,the fused feature was sent to the Transformer encoder to obtain the global feature dependency and sent to the backbone network to extract the global-local fusion feature.At the same time,the two-way attention module was used to enhance the fusion effect of the two branch features.Through experiments on five public data sets,the results show that the strategy network has achieved good performance in four evaluation indicators.

关 键 词：RGB-D显著性检测 全局-局部特征 跨模态融合 多尺度 TRANSFORMER 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]